{"id":30606,"date":"2025-04-22T10:20:58","date_gmt":"2025-04-22T02:20:58","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=30606"},"modified":"2025-04-22T10:20:58","modified_gmt":"2025-04-22T02:20:58","slug":"google-fabu-gemma","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/ja\/google-fabu-gemma\/","title":{"rendered":"Google \u53d1\u5e03 Gemma 3 QAT \u6a21\u578b\uff1a\u8ba9\u9876\u5c16 AI \u6a21\u578b\u8d70\u8fdb\u6d88\u8d39\u7ea7 GPU"},"content":{"rendered":"<p>\u7ee7\u4e0a\u6708\u63a8\u51fa\u65b0\u4e00\u4ee3\u5f00\u653e\u6a21\u578b <a href=\"https:\/\/www.kdjingpai.com\/en\/gemma-3-jishubaogao\/\">Gemma 3<\/a> \u540e\uff0cGoogle \u8fd1\u65e5\u518d\u6b21\u62d3\u5c55\u5176\u6a21\u578b\u751f\u6001\uff0c\u53d1\u5e03\u4e86\u57fa\u4e8e\u91cf\u5316\u611f\u77e5\u8bad\u7ec3 (Quantization-Aware Training, QAT) \u4f18\u5316\u7684\u65b0\u7248\u672c\u3002 Gemma 3 \u6b64\u524d\u5df2\u51ed\u501f\u5176\u5728\u00a0<code>BF16<\/code>\u00a0\u7cbe\u5ea6\u4e0b\u4e8e\u5355\u4e2a\u9ad8\u7aef GPU\uff08\u5982\u00a0<code>NVIDIA H100<\/code>\uff09\u4e0a\u8fd0\u884c\u7684\u80fd\u529b\uff0c\u8bc1\u660e\u4e86\u5176\u9886\u5148\u6027\u80fd\u3002\u800c QAT \u7248\u672c\u7684\u63a8\u51fa\uff0c\u65e8\u5728\u663e\u8457\u964d\u4f4e\u6a21\u578b\u7684\u5185\u5b58\u9700\u6c42\uff0c\u540c\u65f6\u5c3d\u53ef\u80fd\u4fdd\u6301\u9ad8\u8d28\u91cf\u8f93\u51fa\uff0c\u4ece\u800c\u8ba9\u5f3a\u5927\u7684 AI \u6a21\u578b\u80fd\u591f\u5728\u66f4\u591a\u6d88\u8d39\u7ea7\u786c\u4ef6\u4e0a\u8fd0\u884c\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-30607\" title=\"Google \u53d1\u5e03 Gemma 3 QAT \u6a21\u578b\uff1a\u8ba9\u9876\u5c16 AI \u6a21\u578b\u8d70\u8fdb\u6d88\u8d39\u7ea7 GPU-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/04\/48bb51408fca2e4.png\" alt=\"Google \u53d1\u5e03 Gemma 3 QAT \u6a21\u578b\uff1a\u8ba9\u9876\u5c16 AI \u6a21\u578b\u8d70\u8fdb\u6d88\u8d39\u7ea7 GPU-1\" width=\"2019\" height=\"1283\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/04\/48bb51408fca2e4.png 2019w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/04\/48bb51408fca2e4-768x488.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/04\/48bb51408fca2e4-1536x976.png 1536w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/04\/48bb51408fca2e4-18x12.png 18w\" sizes=\"auto, (max-width: 2019px) 100vw, 2019px\" \/><br \/>\n<em>\u8be5\u56fe\u8868\u4f9d\u636e Chatbot Arena Elo \u5206\u6570\u5bf9 AI \u6a21\u578b\u8fdb\u884c\u6392\u540d\uff0c\u5206\u6570\u8d8a\u9ad8\u4ee3\u8868\u7528\u6237\u504f\u597d\u5ea6\u8d8a\u9ad8\u3002\u5706\u70b9\u6807\u6ce8\u4e86\u6a21\u578b\u8fd0\u884c\u6240\u9700\u00a0<code>NVIDIA H100<\/code>\u00a0GPU \u7684\u9884\u4f30\u6570\u91cf\u3002<\/em><\/p>\n<p>&nbsp;<\/p>\n<h3>\u7406\u89e3\u6027\u80fd\u3001\u7cbe\u5ea6\u4e0e\u91cf\u5316\u95e8\u69db<\/h3>\n<p>\u4e0a\u56fe\u5c55\u793a\u4e86\u8fd1\u671f\u53d1\u5e03\u7684\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u6027\u80fd\uff08Elo \u5206\u6570\uff09\u5bf9\u6bd4\u3002\u66f4\u9ad8\u7684\u67f1\u72b6\u56fe\u610f\u5473\u7740\u6a21\u578b\u5728\u4eba\u7c7b\u8bc4\u4f30\u8005\u8fdb\u884c\u7684\u533f\u540d\u5e76\u6392\u6bd4\u8f83\u4e2d\u8868\u73b0\u66f4\u4f73\u3002\u56fe\u8868\u4e0b\u65b9\u6807\u6ce8\u4e86\u5728\u00a0<code>BF16<\/code>\u00a0\u6570\u636e\u7c7b\u578b\u4e0b\u8fd0\u884c\u5404\u6a21\u578b\u6240\u9700\u7684\u00a0<code>NVIDIA H100<\/code>\u00a0GPU \u9884\u4f30\u6570\u91cf\u3002<\/p>\n<p>\u9009\u62e9\u00a0<code>BF16<\/code>\u00a0\u8fdb\u884c\u6bd4\u8f83\uff0c\u662f\u56e0\u4e3a\u5b83\u662f\u8bb8\u591a\u5927\u578b\u6a21\u578b\u63a8\u7406\u65f6\u5e38\u7528\u7684\u6570\u503c\u683c\u5f0f\uff0c\u4ee3\u8868\u6a21\u578b\u53c2\u6570\u4ee5 16 \u4f4d\u7cbe\u5ea6\u5b58\u50a8\u3002\u5728\u7edf\u4e00\u7684\u00a0<code>BF16<\/code>\u00a0\u8bbe\u7f6e\u4e0b\u6bd4\u8f83\uff0c\u6709\u52a9\u4e8e\u5728\u76f8\u540c\u7684\u63a8\u7406\u914d\u7f6e\u4e2d\u8bc4\u4f30\u6a21\u578b\u672c\u8eab\u7684\u56fa\u6709\u80fd\u529b\uff0c\u6392\u9664\u4e86\u4e0d\u540c\u786c\u4ef6\u6216\u91cf\u5316\u7b49\u4f18\u5316\u6280\u672f\u5e26\u6765\u7684\u53d8\u91cf\u3002<\/p>\n<p>\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u867d\u7136\u8be5\u56fe\u8868\u4f7f\u7528\u00a0<code>BF16<\/code>\u00a0\u8fdb\u884c\u516c\u5e73\u6bd4\u8f83\uff0c\u4f46\u5728\u5b9e\u9645\u90e8\u7f72\u8d85\u5927\u578b\u6a21\u578b\u65f6\uff0c\u4e3a\u4e86\u964d\u4f4e\u5bf9\u786c\u4ef6\uff08\u5982\u56fe\u5f62\u5904\u7406\u5668\u6570\u91cf\uff09\u7684\u5de8\u5927\u9700\u6c42\uff0c\u901a\u5e38\u9700\u8981\u91c7\u7528\u5982\u00a0<code>FP8<\/code>\u00a0\u7b49\u66f4\u4f4e\u7cbe\u5ea6\u7684\u683c\u5f0f\uff0c\u8fd9\u53ef\u80fd\u9700\u8981\u5728\u6027\u80fd\u548c\u53ef\u884c\u6027\u4e4b\u95f4\u505a\u51fa\u6743\u8861\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>\u6253\u7834\u786c\u4ef6\u58c1\u5792\uff1a\u8ba9 AI \u66f4\u6613\u83b7\u5f97<\/h3>\n<p>\u5c3d\u7ba1 Gemma 3 \u5728\u9ad8\u7aef\u786c\u4ef6\u4e0a\u8868\u73b0\u51fa\u8272\uff0c\u9002\u7528\u4e8e\u4e91\u90e8\u7f72\u548c\u7814\u7a76\u573a\u666f\uff0c\u4f46\u7528\u6237\u53cd\u9988\u6e05\u6670\u5730\u8868\u660e\uff0c\u5e02\u573a\u671f\u671b\u80fd\u5728\u73b0\u6709\u786c\u4ef6\u4e0a\u8fd0\u7528 Gemma 3 \u7684\u80fd\u529b\u3002\u63a8\u52a8\u5f3a\u5927 AI \u6280\u672f\u7684\u666e\u53ca\uff0c\u610f\u5473\u7740\u9700\u8981\u8ba9\u6a21\u578b\u5728\u53f0\u5f0f\u673a\u3001\u7b14\u8bb0\u672c\u7535\u8111\u751a\u81f3\u79fb\u52a8\u8bbe\u5907\u4e2d\u5e38\u89c1\u7684\u6d88\u8d39\u7ea7 GPU \u4e0a\u4e5f\u80fd\u9ad8\u6548\u8fd0\u884c\u3002\u5728\u672c\u5730\u8fd0\u884c\u5927\u578b\u6a21\u578b\uff0c\u4e0d\u4ec5\u53ef\u4ee5\u589e\u5f3a\u6570\u636e\u9690\u79c1\u6027\u3001\u964d\u4f4e\u7f51\u7edc\u5ef6\u8fdf\u3001\u652f\u6301\u79bb\u7ebf\u5e94\u7528\uff0c\u8fd8\u4e3a\u7528\u6237\u63d0\u4f9b\u4e86\u66f4\u5927\u7684\u5b9a\u5236\u5316\u7a7a\u95f4\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>QAT \u6280\u672f\uff1a\u5e73\u8861\u6027\u80fd\u4e0e\u53ef\u8bbf\u95ee\u6027\u7684\u5173\u952e<\/h3>\n<p>\u91cf\u5316\u6280\u672f\u662f\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807\u7684\u5173\u952e\u3002\u5728 AI \u6a21\u578b\u4e2d\uff0c\u91cf\u5316\u65e8\u5728\u964d\u4f4e\u6a21\u578b\u5b58\u50a8\u548c\u8ba1\u7b97\u65f6\u4f7f\u7528\u7684\u6570\u5b57\uff08\u5373\u6a21\u578b\u53c2\u6570\uff09\u7684\u7cbe\u5ea6\u3002\u8fd9\u7c7b\u4f3c\u4e8e\u901a\u8fc7\u51cf\u5c11\u56fe\u50cf\u4f7f\u7528\u7684\u989c\u8272\u6570\u91cf\u6765\u538b\u7f29\u56fe\u50cf\u3002\u76f8\u8f83\u4e8e\u6bcf\u4e2a\u6570\u5b57\u4f7f\u7528 16 \u4f4d (<code>BF16<\/code>)\uff0c\u91cf\u5316\u53ef\u4ee5\u4f7f\u7528\u66f4\u5c11\u7684\u4f4d\u6570\uff0c\u4f8b\u5982 8 \u4f4d (<code>int8<\/code>) \u751a\u81f3 4 \u4f4d (<code>int4<\/code>)\u3002<\/p>\n<p>\u91c7\u7528\u00a0<code>int4<\/code>\u00a0\u610f\u5473\u7740\u6bcf\u4e2a\u6570\u5b57\u4ec5\u7528 4 \u4f4d\u8868\u793a\uff0c\u6570\u636e\u5927\u5c0f\u76f8\u6bd4\u00a0<code>BF16<\/code>\u00a0\u51cf\u5c11\u4e86 4 \u500d\u3002\u4f20\u7edf\u7684\u91cf\u5316\u65b9\u6cd5\uff08\u5982\u8bad\u7ec3\u540e\u91cf\u5316 PTQ\uff09\u901a\u5e38\u5728\u6a21\u578b\u8bad\u7ec3\u5b8c\u6210\u540e\u8fdb\u884c\uff0c\u867d\u7136\u6613\u4e8e\u5b9e\u65bd\uff0c\u4f46\u53ef\u80fd\u5bfc\u81f4\u660e\u663e\u7684\u6027\u80fd\u4e0b\u964d\u3002\u6b64\u6b21 Google \u91c7\u7528\u7684 QAT \u6280\u672f\u5219\u6709\u6240\u4e0d\u540c\u3002<\/p>\n<p>QAT \u5728\u6a21\u578b\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5c31\u5f15\u5165\u4e86\u91cf\u5316\u6b65\u9aa4\uff0c\u6a21\u62df\u4f4e\u7cbe\u5ea6\u8fd0\u7b97\u3002\u8fd9\u79cd\u65b9\u6cd5\u5141\u8bb8\u6a21\u578b\u5728\u8bad\u7ec3\u65f6\u5c31\u9002\u5e94\u4f4e\u7cbe\u5ea6\u73af\u5883\uff0c\u4ece\u800c\u5728\u540e\u7eed\u91cf\u5316\u65f6\u80fd\u591f\u66f4\u597d\u5730\u4fdd\u6301\u51c6\u786e\u6027\uff0c\u5b9e\u73b0\u66f4\u5c0f\u3001\u66f4\u5feb\u7684\u6a21\u578b\u3002\u5177\u4f53\u800c\u8a00\uff0cGoogle \u5bf9 Gemma 3 \u6a21\u578b\u5e94\u7528\u4e86\u7ea6 5000 \u6b65\u7684 QAT\uff0c\u4ee5\u672a\u91cf\u5316\u68c0\u67e5\u70b9\u7684\u6982\u7387\u4f5c\u4e3a\u76ee\u6807\u3002\u636e\u79f0\uff0c\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0c\u5728\u91cf\u5316\u5230\u00a0<code>Q4_0<\/code>\u00a0\u683c\u5f0f\u65f6\uff0c\u56f0\u60d1\u5ea6\uff08Perplexity\uff0c\u8861\u91cf\u6a21\u578b\u9884\u6d4b\u80fd\u529b\u7684\u6307\u6807\uff09\u7684\u4e0b\u964d\u5e45\u5ea6\u51cf\u5c11\u4e86 54% \uff08\u4f7f\u7528 <a href=\"https:\/\/www.kdjingpai.com\/en\/llamacpp\/\">llama.cpp<\/a> \u8bc4\u4f30\uff09\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>\u663e\u8457\u6548\u679c\uff1aVRAM \u5360\u7528\u5927\u5e45\u964d\u4f4e<\/h3>\n<p><code>int4<\/code>\u00a0\u91cf\u5316\u5e26\u6765\u7684\u5f71\u54cd\u662f\u663e\u8457\u7684\uff0c\u5c24\u5176\u662f\u5728\u52a0\u8f7d\u6a21\u578b\u6743\u91cd\u6240\u9700\u7684\u663e\u5b58 (VRAM) \u65b9\u9762\uff1a<\/p>\n<ul>\n<li><strong>Gemma 3 27B:<\/strong>\u00a0\u4ece 54 GB (<code>BF16<\/code>) \u964d\u81f3\u00a0<strong>14.1 GB<\/strong>\u00a0(<code>int4<\/code>)<\/li>\n<li><strong>Gemma 3 12B:<\/strong>\u00a0\u4ece 24 GB (<code>BF16<\/code>) \u7f29\u51cf\u81f3\u00a0<strong>6.6 GB<\/strong>\u00a0(<code>int4<\/code>)<\/li>\n<li><strong>Gemma 3 4B:<\/strong>\u00a0\u4ece 8 GB (<code>BF16<\/code>) \u51cf\u5c11\u81f3\u00a0<strong>2.6 GB<\/strong>\u00a0(<code>int4<\/code>)<\/li>\n<li><strong>Gemma 3 1B:<\/strong>\u00a0\u4ece 2 GB (<code>BF16<\/code>) \u964d\u81f3\u00a0<strong>0.5 GB<\/strong>\u00a0(<code>int4<\/code>)<\/li>\n<\/ul>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-30608\" title=\"Google \u53d1\u5e03 Gemma 3 QAT \u6a21\u578b\uff1a\u8ba9\u9876\u5c16 AI \u6a21\u578b\u8d70\u8fdb\u6d88\u8d39\u7ea7 GPU-2\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/04\/5a32673489f6a59.png\" alt=\"Google \u53d1\u5e03 Gemma 3 QAT \u6a21\u578b\uff1a\u8ba9\u9876\u5c16 AI \u6a21\u578b\u8d70\u8fdb\u6d88\u8d39\u7ea7 GPU-2\" width=\"2019\" height=\"1410\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/04\/5a32673489f6a59.png 2019w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/04\/5a32673489f6a59-768x536.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/04\/5a32673489f6a59-1536x1073.png 1536w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/04\/5a32673489f6a59-18x12.png 18w\" sizes=\"auto, (max-width: 2019px) 100vw, 2019px\" \/><\/p>\n<p><strong>\u6ce8\u610f:<\/strong>\u00a0<em>\u6b64\u6570\u636e\u4ec5\u4ee3\u8868\u52a0\u8f7d\u6a21\u578b\u6743\u91cd\u6240\u9700\u7684 VRAM\u3002\u8fd0\u884c\u6a21\u578b\u8fd8\u9700\u8981\u989d\u5916\u7684 VRAM \u7528\u4e8e KV \u7f13\u5b58\uff0c\u8be5\u7f13\u5b58\u5b58\u50a8\u5bf9\u8bdd\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u5176\u5927\u5c0f\u53d6\u51b3\u4e8e\u4e0a\u4e0b\u6587\u957f\u5ea6\u3002<\/em><\/p>\n<p>&nbsp;<\/p>\n<h3>\u5728\u4f60\u7684\u8bbe\u5907\u4e0a\u8fd0\u884c Gemma 3<\/h3>\n<p>\u663e\u5b58\u9700\u6c42\u7684\u5927\u5e45\u964d\u4f4e\uff0c\u4f7f\u5f97\u5728\u5e7f\u6cdb\u4f7f\u7528\u7684\u6d88\u8d39\u7ea7\u786c\u4ef6\u4e0a\u8fd0\u884c\u66f4\u5927\u3001\u66f4\u5f3a\u7684\u6a21\u578b\u6210\u4e3a\u53ef\u80fd\uff1a<\/p>\n<ul>\n<li><strong>Gemma 3 27B (<code>int4<\/code>):<\/strong>\u00a0\u73b0\u5728\u53ef\u4ee5\u8f7b\u677e\u8fd0\u884c\u5728\u5355\u4e2a\u62e5\u6709 24GB VRAM \u7684\u684c\u9762\u7ea7\u00a0<code>NVIDIA RTX 3090<\/code>\u00a0\u6216\u7c7b\u4f3c\u663e\u5361\u4e0a\uff0c\u7528\u6237\u5f97\u4ee5\u5728\u672c\u5730\u8fd0\u884c Google \u6700\u5927\u7684 Gemma 3 \u53d8\u4f53\u3002<\/li>\n<li><strong>Gemma 3 12B (<code>int4<\/code>):<\/strong>\u00a0\u80fd\u591f\u5728\u62e5\u6709 8GB VRAM \u7684\u7b14\u8bb0\u672c GPU\uff08\u5982\u00a0<code>NVIDIA RTX 4060 Laptop GPU<\/code>\uff09\u4e0a\u9ad8\u6548\u8fd0\u884c\uff0c\u5c06\u5f3a\u5927\u7684 AI \u80fd\u529b\u5e26\u5230\u4fbf\u643a\u8bbe\u5907\u3002<\/li>\n<li><strong>\u66f4\u5c0f\u6a21\u578b (4B, 1B):<\/strong>\u00a0\u4e3a\u8d44\u6e90\u66f4\u53d7\u9650\u7684\u7cfb\u7edf\uff08\u5305\u62ec\u90e8\u5206\u79fb\u52a8\u8bbe\u5907\uff09\u63d0\u4f9b\u4e86\u66f4\u9ad8\u7684\u53ef\u8bbf\u95ee\u6027\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h3>\u4e0e\u4e3b\u6d41\u5de5\u5177\u8f7b\u677e\u96c6\u6210<\/h3>\n<p>\u4e3a\u4e86\u65b9\u4fbf\u7528\u6237\u5728\u81ea\u5df1\u7684\u5de5\u4f5c\u6d41\u4e2d\u4f7f\u7528\u8fd9\u4e9b\u6a21\u578b\uff0cGoogle \u5df2\u5728\u00a0<code>Hugging Face<\/code>\u00a0\u548c\u00a0<code>Kaggle<\/code>\u00a0\u4e0a\u63d0\u4f9b\u4e86\u5b98\u65b9\u7684\u00a0<code>int4<\/code>\u00a0\u548c\u00a0<code>Q4_0<\/code>\u00a0\u672a\u91cf\u5316 QAT \u6a21\u578b\u3002\u540c\u65f6\uff0cGoogle \u4e0e\u591a\u4e2a\u6d41\u884c\u7684\u5f00\u53d1\u8005\u5de5\u5177\u8fdb\u884c\u4e86\u5408\u4f5c\uff0c\u4ee5\u4fbf\u7528\u6237\u65e0\u7f1d\u8bd5\u7528\u57fa\u4e8e QAT \u7684\u91cf\u5316\u68c0\u67e5\u70b9\uff1a<\/p>\n<ul>\n<li><strong>Ollama:<\/strong>\u00a0\u901a\u8fc7\u7b80\u5355\u547d\u4ee4\u5373\u53ef\u5feb\u901f\u8fd0\u884c\uff0c\u6240\u6709 Gemma 3 QAT \u6a21\u578b\u5df2\u83b7\u5f97\u539f\u751f\u652f\u6301\u3002<\/li>\n<li><strong>LM Studio:<\/strong>\u00a0\u63d0\u4f9b\u7528\u6237\u53cb\u597d\u7684\u754c\u9762\uff0c\u53ef\u5728\u684c\u9762\u4e0a\u8f7b\u677e\u4e0b\u8f7d\u548c\u8fd0\u884c Gemma 3 QAT \u6a21\u578b\u3002<\/li>\n<li><strong>MLX:<\/strong>\u00a0\u5229\u7528\u00a0<code>MLX<\/code>\u00a0\u6846\u67b6\u5728\u00a0<code>Apple Silicon<\/code>\u00a0\u4e0a\u5b9e\u73b0 Gemma 3 QAT \u6a21\u578b\u7684\u9ad8\u6548\u3001\u4f18\u5316\u63a8\u7406\u3002<\/li>\n<li><strong>Gemma.cpp:<\/strong>\u00a0Google \u63d0\u4f9b\u7684\u4e13\u7528 C++ \u5b9e\u73b0\uff0c\u53ef\u5728 CPU \u4e0a\u8fdb\u884c\u9ad8\u6548\u63a8\u7406\u3002<\/li>\n<li><strong>llama.cpp:<\/strong>\u00a0\u5bf9 GGUF \u683c\u5f0f\u7684 QAT \u6a21\u578b\u63d0\u4f9b\u539f\u751f\u652f\u6301\uff0c\u4fbf\u4e8e\u96c6\u6210\u5230\u73b0\u6709\u5de5\u4f5c\u6d41\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h3>Gemmaverse \u751f\u6001\u4e2d\u7684\u66f4\u591a\u91cf\u5316\u9009\u62e9<\/h3>\n<p>\u9664\u4e86\u5b98\u65b9\u53d1\u5e03\u7684 QAT \u6a21\u578b\uff0c\u6d3b\u8dc3\u7684\u00a0<code>Gemmaverse<\/code>\u00a0\u793e\u533a\u4e5f\u63d0\u4f9b\u4e86\u8bb8\u591a\u66ff\u4ee3\u65b9\u6848\u3002\u8fd9\u4e9b\u65b9\u6848\u901a\u5e38\u91c7\u7528\u8bad\u7ec3\u540e\u91cf\u5316 (PTQ) \u6280\u672f\uff0c\u7531\u00a0<code>Bartowski<\/code>,\u00a0<code><a href=\"https:\/\/www.kdjingpai.com\/en\/unsloth\/\">Unsloth<\/a><\/code>,\u00a0<code>GGML<\/code>\u00a0\u7b49\u793e\u533a\u6210\u5458\u8d21\u732e\uff0c\u5e76\u5728\u00a0<code>Hugging Face<\/code>\u00a0\u4e0a\u63d0\u4f9b\u3002\u63a2\u7d22\u8fd9\u4e9b\u793e\u533a\u9009\u9879\u4e3a\u7528\u6237\u63d0\u4f9b\u4e86\u5728\u6a21\u578b\u5927\u5c0f\u3001\u901f\u5ea6\u548c\u8d28\u91cf\u4e4b\u95f4\u8fdb\u884c\u6743\u8861\u7684\u66f4\u5e7f\u6cdb\u9009\u62e9\uff0c\u4ee5\u6ee1\u8db3\u7279\u5b9a\u9700\u6c42\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>\u5373\u523b\u5f00\u59cb\u4f53\u9a8c<\/h3>\n<p>\u5c06\u524d\u6cbf\u7684 AI \u6027\u80fd\u5e26\u5230\u6613\u4e8e\u83b7\u53d6\u7684\u786c\u4ef6\u4e0a\uff0c\u662f\u63a8\u52a8 AI \u5f00\u53d1\u6c11\u4e3b\u5316\u7684\u5173\u952e\u4e00\u6b65\u3002\u501f\u52a9\u901a\u8fc7 QAT \u4f18\u5316\u7684 Gemma 3 \u6a21\u578b\uff0c\u7528\u6237\u73b0\u5728\u53ef\u4ee5\u5728\u81ea\u5df1\u7684\u53f0\u5f0f\u673a\u6216\u7b14\u8bb0\u672c\u7535\u8111\u4e0a\u5229\u7528\u5c16\u7aef AI \u80fd\u529b\u3002\u8fd9\u4e0d\u4ec5\u964d\u4f4e\u4e86\u5f00\u53d1\u8005\u548c\u7814\u7a76\u4eba\u5458\u63a2\u7d22\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u95e8\u69db\uff0c\u4e5f\u9884\u793a\u7740\u66f4\u5f3a\u5927\u7684\u8fb9\u7f18 AI \u5e94\u7528\u7684\u6f5c\u529b\u3002<\/p>\n<p>\u63a2\u7d22\u91cf\u5316\u6a21\u578b\u5e76\u5f00\u59cb\u6784\u5efa\uff1a<\/p>\n<ul>\n<li>\u5728 PC \u4e0a\u4f7f\u7528\u00a0<code><a href=\"https:\/\/www.kdjingpai.com\/en\/ollama\/\">Ollama<\/a><\/code>\u00a0\u8fd0\u884c\uff1a<a href=\"https:\/\/ollama.com\/library\/gemma3\">https:\/\/ollama.com\/library\/gemma3<\/a><\/li>\n<li>\u5728\u00a0<code>Hugging Face<\/code>\u00a0\u548c\u00a0<code>Kaggle<\/code>\u00a0\u4e0a\u67e5\u627e\u6a21\u578b\uff1a\n<ul>\n<li>Hugging Face:\u00a0<a href=\"https:\/\/huggingface.co\/collections\/google\/gemma-3-qat-67ee61ccacbf2be4195c265b\">https:\/\/huggingface.co\/collections\/google\/gemma-3-qat-67ee61ccacbf2be4195c265b<\/a><\/li>\n<li>Kaggle:\u00a0<a href=\"https:\/\/www.kaggle.com\/models\/google\/gemma-3\/transformers\">https:\/\/www.kaggle.com\/models\/google\/gemma-3\/transformers<\/a><\/li>\n<\/ul>\n<\/li>\n<li>\u5728\u79fb\u52a8\u8bbe\u5907\u4e0a\u8fd0\u884c\uff08\u901a\u8fc7\u00a0<code>Google AI Edge<\/code>\uff09\uff1a<a href=\"https:\/\/developers.googleblog.com\/en\/gemma-3-on-mobile-and-web-with-google-ai-edge\/\">https:\/\/developers.googleblog.com\/en\/gemma-3-on-mobile-and-web-with-google-ai-edge\/<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\u7ee7\u4e0a\u6708\u63a8\u51fa\u65b0\u4e00\u4ee3\u5f00\u653e\u6a21\u578b Gemma 3 \u540e\uff0cGoogle \u8fd1\u65e5\u518d\u6b21\u62d3\u5c55\u5176\u6a21\u578b\u751f\u6001\uff0c\u53d1\u5e03\u4e86\u57fa\u4e8e\u91cf\u5316\u611f\u77e5\u8bad\u7ec3 (Quantization-Aware Training, QAT) \u4f18\u5316\u7684\u65b0\u7248\u672c\u3002 Gemma 3 \u6b64\u524d\u5df2\u51ed\u501f\u5176\u5728\u00a0BF16\u00a0\u7cbe\u5ea6&#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[46],"tags":[],"class_list":["post-30606","post","type-post","status-publish","format-standard","hentry","category-news"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/30606","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/comments?post=30606"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/30606\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/media?parent=30606"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/categories?post=30606"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/tags?post=30606"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}